Add the major Falcon Moddels 7b and 40b instruct #37
Conversation
Change schema.json to reflect falcon 40B models
|
Sorry for the delay. Thanks for this! Will come back here after the next version comes out (hopefully tomorrow) |
| "datePublished": "2023-10-22T03:04:42", | ||
| "name": "falcon-40b-instruct", | ||
| "description": "Falcon-40B-Instruct, based on Falcon-40B, has been fine-tuned on chat and instruct datasets. It offers outstanding performance and is an excellent choice for chat and instruct applications. Falcon-7B-Instruct is part of the Falcon family of language models, known for their exceptional capabilities and openness.", | ||
| "author": { |
There was a problem hiding this comment.
The author field is intended for the model creator, so in this case it would talk about tiiuae. See other files for example
| "url": "https://huggingface.co/maddes8cht", | ||
| "blurb": "Maddes8cht Passionate about Open Source and AI. On Hugging Face he is advocating for real open source AI models with OSI compliant licenses" | ||
| }, | ||
| "numParameters": "40B", |
There was a problem hiding this comment.
let's actually remove the 40B model from here: people might not realize the resources this model will require and they won't understand why it's not working. As a rule of thumb, catalog models should be 13B and below
| "_descriptorVersion": "0.0.1", | ||
| "datePublished": "2023-10-31T16:01:50", | ||
| "name": "falcon-7b-instruct", | ||
| "description": "Falcon-7B-Instruct, based on Falcon-7B, has been fine-tuned on chat and instruct datasets. It offers outstanding performance and is an excellent choice for chat and instruct applications. Falcon-7B-Instruct is part of the Falcon family of language models, known for their exceptional capabilities and openness.", |
There was a problem hiding this comment.
I typically generate description by following this process, more or less:
- copy the parts from the original model card with information about the model, the dataset, training process, prompting nuances, license
- feed this information to GPT4 along with a prompt that looks something like:
Please generate a short summary of this information. The audience is AI model users with varying technical backgrounds, from non-technical users to ML researchers. The summary must include every important detail in the text above. Use a scientific and dispassionate tone.
feel free to tweak it
| "files": { | ||
| "highlighted": { | ||
| "economical": { | ||
| "name": "tiiuae-falcon-7b-instruct-Q4_K_M" |
There was a problem hiding this comment.
these need to be the same as the filename to be downloaded
| "name": "tiiuae-falcon-7b-instruct-Q2_K.gguf", | ||
| "url": "https://huggingface.co/maddes8cht/tiiuae-falcon-7b-instruct-gguf/blob/main/tiiuae-falcon-7b-instruct-Q2_K.gguf", | ||
| "sizeBytes": 4025162688, | ||
| "quantization": "Q2_K", |
There was a problem hiding this comment.
Economical should be Q4_K_S in most cases (also it should match the filename under "files" above)
| "repositoryUrl": "https://huggingface.co/maddes8cht/tiiuae-falcon-7b-instruct-gguf" | ||
| }, | ||
| { | ||
| "name": "tiiuae-falcon-7b-instruct-Q4_K_M.gguf", |
There was a problem hiding this comment.
No need to include files here that aren't in the highlighted section. The max should be 2
Add the major Falcon Moddels, 7b instruct and 40b instruct..
Change schema.json to reflect falcon 40B models:
In schema.json there is one quantization type missing in
definition - ModelFile-properties: There is a
Q3_K_Ltype, i added it to the enum.The apache-licensed Falcon Models come in the flavours of 7B and 40B.
A 40B option is missing in the enumeration
numParameters, so i added "40B" to the enum.I have a bunch of other interesting and powerful Falcon-based Models, notably the wizardLM uncensored versions, that i would like to ad to this list.
I will do so if you accept these changes.